SAN JOSE, Calif. Cisco Systems has designed a set of proprietary ASICs to more than double the DRAM memory linked to an Intel Nehalem processor. The company aims to use the technology to leapfrog existing server makers in areas such as database performance or the number of virtual machines a server can support.
Cisco announced its plan to get into the server business in March, competing with the likes of IBM and Hewlett-Packard. At the time it said the close integration of its Fibre Channel over Ethernet technology with its servers would be a major differentiating factor along with the Nehalem memory expansion.
"If you are going to be the newcomer to the market you need to have something to get customers to take a look at you," said Nathan Brookwood, principal of market watcher Insight64 (Saratoga, Calif.). "I don't know without more details precisely how much advantage they will achieve with this approach," he said, adding he had not yet been briefed on Cisco's memory technology.
The Cisco Unified Computing System includes a computer blade—the B250M1--that can support up to 48 dual in-line memory modules hosting up to 384 Gbytes RAM using 8 Gbyte DIMMs. By contrast, other Nehalem servers support either 12 or 18 DIMMs and up to 144 Gbytes DRAM.
Cisco designed a set of ASICs that fan out connections between the memory controller integrated in the Nehalem and DIMM slots on the board. The approach uses at least two separate ASIC designs and at least six chips per blade.
The ASICs manage the memory channel and rank structures Intel defines in the Nehalem memory controller. Cisco has not yet determined whether it will run the memory bus at the full 1,066 MHz data rate or at a lower 800 MHz rate.
Intel provides a reference design for two-socket Nehalem servers that support 12 DIMMs at 1,066 MHz and another that supports 18 DIMMs at 800 MHz.
Cisco is not providing any other details on the ASICs. The UCS systems will ship in June but the B250M1 card with the memory expansion ASICs will not ship until the fall.
"Clearly if you add more memory to a Nehalem processor you could be increasing the latency on the memory and that could impact performance," Brookwood said. "At some point you get out of balance in memory and processing power and the trick is to be in balance," he said.
The approach also adds costs in the form of the ASICs and DRAMs, he added. "This could me a very expensive alternative," Brookwood said.
Cisco has not published the cost of the server boards yet. The approach "will definitely save a lot of money" compared to adding more servers to add performance, said Brian Schwarz, a principal engineer in Cisco's UCS product management team.
The company has not determined how many additional virtual machines the servers can support with the memory expansion. Exact figures will probably vary based on the specific workload, Schwarz said.
Similarly, systems could get a performance boost by putting more of their databases in RAM on the card. But the amount of performance boost will depend on the specifics of the user's database.
Cisco is one of the world's largest developers of ASICs, used liberally in its routers and switches. Most computer servers use only off-the-shelf chips to save costs. Indeed, HP is said to have a policy forbidding use of ASICs in its computer systems.
The memory expansion chips are the only new ASICs in Cisco's UCS systems. The system does include other ASICs repurposed from Cisco's Nexus line of Fibre Channel over Ethernet switches.
Each UCS chassis uses two Nexus ASICs as fabric extenders. Each chip supports up to 10 10Gbit/second Ethernet links. In the Nexus systems they are configured to support four 10G and 48 Gbit links.